1. Bibliographic Information

1.1. Title

The title of the paper is "SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model". The central topic is the development of a comprehensive framework and dataset for detecting, localizing, and explaining deepfakes in social media images.

1.2. Authors

The authors of the paper are:

• Zhenglin Huang (University of Liverpool, UK)

• Jinwei Hu (University of Liverpool, UK)

• Xiangtai Li (Nanyang Technological University, SG)

• Yiwei He (University of Liverpool, UK)

• Xingyu Zhao (WMG, University of Warwick)

• Bei Peng (University of Liverpool, UK)

• Baoyuan Wu (The Chinese University of Hong Kong, Shenzhen, Guangdong, China)

• Xiaowei Huang (University of Liverpool, UK)

• Guangliang Cheng (University of Liverpool, UK)

  The corresponding authors are Xiangtai Li and Guangliang Cheng. The authors are primarily affiliated with academic institutions in the UK, Singapore, and China, indicating a collaborative international research effort in the fields of computer vision, artificial intelligence, and deepfake detection.

1.3. Journal/Conference

While the paper's original source link is a local file path (/files/papers/.../paper.pdf), indicating it might be a preprint, it is presented as a research paper. Such papers are typically submitted to prestigious conferences or journals in computer vision and machine learning (e.g., CVPR, ICCV, NeurIPS, ICML, IEEE TPAMI). The technical depth and comparative analysis suggest it is intended for a top-tier venue. Often, papers first appear on arXiv before formal publication.

1.4. Publication Year

The publication year is implied to be 2024, as indicated by the copyright information in the abstract ("2024").

1.5. Abstract

The paper addresses the growing threat of realistic AI-generated images contributing to misinformation, particularly on social media. It highlights the current lack of large, diverse deepfake detection datasets tailored for social media and effective solutions. To counter this, the authors introduce SID-Set (Social media Image Detection dataSet), a novel dataset comprising 300K AI-generated/tampered and authentic images with extensive annotations. SID-Set is characterized by its volume, diversity (covering fully synthetic and tampered images), and elevated realism. Building on this, they propose SIDA (Social media Image Detection, localization, and explanation Assistant), a framework utilizing large multimodal models. SIDA not only classifies image authenticity but also pinpoints tampered regions via mask prediction and offers textual explanations for its judgments. Extensive experiments demonstrate SIDA's superior performance compared to state-of-the-art models on SID-Set and other benchmarks. The code, model, and dataset are planned for release.

1.6. Original Source Link

The original source link provided is: /files/papers/69128938b150195a0db74a7e/paper.pdf. This appears to be a local file path, suggesting the paper was accessed from a local directory or internal system. Its publication status (e.g., officially published in a journal/conference or a preprint on arXiv) is not explicitly stated in the provided text.

2. Executive Summary

2.1. Background & Motivation

The core problem the paper aims to solve is the proliferation of highly realistic AI-generated images (deepfakes) on social media, which poses substantial risks for misinformation dissemination. These synthetic images can mislead vast audiences, erode trust in digital content, and have severe real-world repercussions.

This problem is critically important in the current digital age due to the rapid advancements in generative models (e.g., Diffusion Models, GANs). These models can create images that are increasingly difficult for humans, and even existing deepfake detection models, to distinguish from genuine ones.

Specific challenges or gaps exist in prior research:

1. Insufficient Dataset Diversity: Most existing deepfake datasets primarily focus on facial imagery, neglecting the broader issue of non-facial image falsification, which is becoming prevalent. While some larger datasets exist for general image deepfake detection, they often use outdated generative techniques or do not specifically target social media content, leading to less convincing forgeries.

2. Limited Dataset Comprehensiveness: Existing datasets are typically designed either for deepfake detection (binary classification) or for tampered region localization, often focusing on specific manipulation types. An ideal dataset should encompass a wide range of scenarios, including fully synthetic and tampered images, reflecting the complexity of real social media content. Furthermore, they lack interpretability, meaning they don't explain why a decision was made.

   The paper's entry point or innovative idea is to address these gaps by:

• Creating SID-Set, a large, diverse, and realistic dataset specifically for social media deepfakes that includes comprehensive annotations for detection, localization, and textual explanations.
• Proposing SIDA, a novel framework that leverages the capabilities of large multimodal models (LMMs) to not only detect deepfakes and localize manipulations but also provide textual explanations for its judgments, enhancing transparency and utility.

2.2. Main Contributions / Findings

The paper's primary contributions are threefold:

1. Establishment of SID-Set:

   • A comprehensive benchmark dataset for detecting, localizing, and explaining deepfakes in social media images.
   • Features an extensive volume of 300K images (100K real, 100K synthetic, 100K tampered).
   • Offers broad diversity, covering both fully synthetic and tampered images across various classes.
   • Achieves elevated realism, with images predominantly indistinguishable from genuine ones through visual inspection.
   • Includes comprehensive annotations, such as masks for tampered regions and textual descriptions explaining judgment criteria.

2. Proposal of SIDA Framework:

   • A new image deepfake detection, localization, and explanation framework built upon large multimodal models (VLMs).
   • Goes beyond traditional binary classification by also delineating tampered regions (mask prediction) and providing textual explanations of the model's judgment criteria.
   • Designed to be a comprehensive solution for tackling the complex nature of social media image deepfakes.

3. Extensive Experimental Validation:

   • Demonstrates that SIDA achieves superior performance compared to state-of-the-art deepfake detection models on the SID-Set and other benchmarks.

   • Effectively identifies and delineates tampered areas within images.

   • Showcases robustness against common image perturbations (e.g., JPEG compression, resizing, Gaussian noise).

     The key conclusions reached by the paper are that SIDA, supported by the SID-Set, offers a more robust, interpretable, and comprehensive approach to deepfake detection in social media. These findings directly address the problems of insufficient dataset diversity, limited comprehensiveness, and lack of interpretability in existing deepfake detection systems.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To fully understand this paper, a novice reader should be familiar with the following concepts:

• Generative Models: These are a class of artificial intelligence models that can generate new data instances that resemble the training data. In the context of images, this means creating realistic images, often from noise or text descriptions. Examples include Generative Adversarial Networks (GANs) and Diffusion Models. The paper specifically mentions Diffusion Models (e.g., FLUX, SDXL) as a key technology for creating realistic synthetic images.
• Deepfake: A deepfake is a synthetic media in which a person in an existing image or video is replaced with someone else's likeness. More broadly, it refers to any AI-generated or manipulated content that is highly realistic and intended to deceive. The paper categorizes deepfakes into fully synthetic images (entirely AI-generated) and tampered images (real images with AI-modified regions).
• Deepfake Detection: The task of identifying whether an image or video is a deepfake or authentic (real). This is often framed as a classification task (real vs. fake).
• Deepfake Localization: Beyond just detecting a deepfake, localization involves identifying the specific regions or pixels within an image that have been manipulated. This is typically achieved by generating a segmentation mask.
• Large Multimodal Models (LMMs) / Vision-Language Models (VLMs): These are advanced AI models that can process and understand information from multiple modalities, most commonly vision (images/videos) and language (text). They learn to align visual features with textual descriptions, enabling tasks like image captioning, visual question answering, and, in this paper's context, deepfake detection and explanation. The paper builds upon models like LLaMA, LLaVA, and LISA.
  • LLaMA (Large Language Model Meta AI): A family of foundational large language models (LLMs) developed by Meta. They are known for their strong language understanding and generation capabilities with relatively compact designs.
  • LLaVA (Large Language and Vision Assistant): A VLM that extends LLaMA by integrating visual instruction tuning. It connects a vision encoder to an LLM to enable multimodal understanding, allowing it to process both images and text and perform tasks like visual question answering.
  • LISA (Reasoning Segmentation via Large Language Model): A VLM that extends LLaVA by incorporating capabilities for fine-grained segmentation. It uses LLMs to understand complex textual instructions and generate precise segmentation masks for objects or regions within an image. LISA is the base model for SIDA.
• Segmentation Mask: A pixel-level annotation that identifies the exact boundary of an object or region within an image. For deepfake localization, a segmentation mask highlights the manipulated pixels.
• Convolutional Neural Networks (CNNs): A class of deep learning models particularly effective for image processing tasks. They use convolutional layers to automatically learn hierarchical features from image data. Many traditional deepfake detectors rely on CNNs.
• Transformers: A neural network architecture that has revolutionized natural language processing and, more recently, computer vision. Key to Transformers is the self-attention mechanism, which allows the model to weigh the importance of different parts of the input data when processing a particular element. The paper mentions Multihead Attention in its methodology.
  • Multihead Attention: An extension of the attention mechanism where multiple "attention heads" learn to attend to different parts of the input independently and in parallel. The outputs from these heads are then concatenated and linearly transformed, allowing the model to capture diverse relationships and dependencies.
• CrossEntropy Loss ($\mathcal{L}_{\mathrm{CE}}$): A common loss function used in classification tasks. It measures the difference between the predicted probability distribution and the true distribution. A lower cross-entropy value indicates a more accurate prediction.
• Binary Cross-Entropy (BCE) Loss ($\mathcal{L}_{\mathrm{BCE}}$): A specific form of cross-entropy loss used for binary classification problems (e.g., pixel is manipulated vs. not manipulated in segmentation).
• DICE Loss ($\mathcal{L}_{\mathrm{DICE}}$): A loss function commonly used in image segmentation tasks, especially when dealing with imbalanced class distributions (e.g., a small manipulated region within a large image). It measures the overlap between the predicted segmentation mask and the ground truth mask.
• LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning technique for large pre-trained models. Instead of fine-tuning all model parameters, LoRA injects small, trainable low-rank matrices into the Transformer layers, significantly reducing the number of trainable parameters and computational cost during fine-tuning while maintaining performance.
• Evaluation Metrics:
  • Accuracy (Acc): The proportion of correctly classified instances (true positives + true negatives) out of the total instances.
  • F1 Score: The harmonic mean of Precision and Recall. It is a good metric when there is an uneven class distribution, as it balances both false positives and false negatives.
  • Area Under the Curve (AUC): In the context of deepfake detection, AUC typically refers to the Area Under the Receiver Operating Characteristic (ROC) curve. The ROC curve plots the True Positive Rate (TPR) against the False Positive Rate (FPR) at various classification thresholds. AUC measures the overall performance of a binary classifier, with a higher value indicating better performance.
  • Intersection over Union (IoU): Also known as the Jaccard Index, IoU is a common metric for evaluating the performance of object detection and segmentation models. It measures the overlap between the predicted bounding box/mask and the ground truth bounding box/mask. A higher IoU indicates better localization accuracy.

3.2. Previous Works

The paper contextualizes its contribution by discussing prior research in three main areas:

3.2.1. Image Deepfake Datasets

Historically, deepfake research concentrated on facial deepfakes due to their immediate societal impact. Key datasets in this domain include:

• ForgeryNet [24]

• DeepFakeFace [58]

• DFFD [7]

  As generative AI advanced, the focus broadened to non-facial deepfake data. Datasets generated using text-to-image or image-to-image techniques (often involving GANs or Stable Diffusion models) emerged:

• GenImage [86]

• HiFiIFDL [21]

• DiffForensics [66]

  These newer datasets aimed for larger volumes, diverse generation methodologies, and richer annotations beyond just real/fake labels, sometimes including segmentation masks for manipulated regions [21, 22, 73].

Background on Dataset Limitations: The paper criticizes existing non-facial deepfake datasets for several reasons:

• Outdated Generative Technologies: Many use older techniques, resulting in lower quality, easily detectable forgeries.
• Limited Scope: They often focus only on text-to-image or image-to-image generation, neglecting localized manipulations of specific objects or parts within an image, which can be more insidious.
• Lack of Authenticity Criteria: They often lack clear criteria for content authenticity or textual explanations, hindering interpretability.
• Focus on Single Deepfake Type: Most emphasize either fully synthetic or tampered images, not a combination, which is common in real social media.

3.2.2. Image Deepfake Detection and Localization

Traditional deepfake detection methods treat the problem as a classification task, using various architectures like CNNs and Transformers to identify artifacts.

• Classification-focused methods: [4, 6, 38, 49, 62] employ strategies like data augmentation [62], adversarial training [6], or reconstruction [4] to improve generality and precision.
• Frequency Domain Analysis: Some researchers extract features from the frequency domain [25, 60] or fuse features from both spatial and frequency domains [13, 65] to obtain more comprehensive discriminative features.
• Localization Methods: A subset of methods goes further by constructing datasets with masks of locally tampered areas to perform both detection and localization [20, 40, 44, 61, 80, 83]. However, these are largely concentrated on facial data, with fewer large, public datasets for non-facial or social media deepfakes.

3.2.3. Large Multimodal Models (LMMs)

The paper notes the significant progress in LLMs [14-16, 41] and VLMs [14-16, 29, 71, 78], which seamlessly integrate visual and textual data.

• LLMs: The LLaMA series [14-16] optimizes language understanding.

• VLMs for Multimodal Understanding: LLaVA series [34, 35] enhances visual question answering by aligning visual and textual data.

• VLMs for Segmentation: LISA series [29, 71] employ LLMs for accurate image segmentation, merging visual perception with linguistic insights.

• Grounding LMMs: Several grounding large multimodal models [50, 54, 55, 67, 68, 72, 74, 77, 81] localize content based on linguistic information.

  Recent advancements have also seen multimodal data integration for deepfake detection:

• AntifakePrompt [5]: Formulates deepfake detection as a visual question answering problem using InstructBLIP.

• ForgeryGPT [31]: Integrates forensic knowledge with a mask-aware forgery extractor for pixel-level fraud detection.

• FakeShield [69]: Leverages LLaVA to identify and localize altered regions while providing interpretable insights. (This is a concurrent work).

3.3. Technological Evolution

The field of deepfake detection has evolved from:

1. Early focus on facial deepfakes: Driven by the immediate impact of manipulated faces, datasets like ForgeryNet emerged, primarily for real/fake classification or facial manipulation localization.
2. Emergence of non-facial deepfakes: With the rise of powerful generative models (GANs, Diffusion Models), the ability to generate or manipulate arbitrary image content improved. This led to datasets like GenImage, but they often suffered from outdated generation methods or lacked comprehensiveness.
3. Integration of multimodal learning: The success of LLMs and VLMs in understanding and generating text, combined with their ability to process images, opened new avenues for deepfake detection. Models like LLaVA and LISA demonstrated multimodal comprehension and fine-grained segmentation.
4. Towards comprehensive and interpretable solutions: The current paper represents a step towards combining detection, localization, and explanation within a single framework, leveraging VLMs and addressing the need for more realistic and diverse social media deepfake datasets. SIDA aims to overcome the limitations of previous methods that were either limited in scope (e.g., facial only), lacked realism, or couldn't provide explanations.

3.4. Differentiation Analysis

Compared to the main methods and datasets in related work, SIDA and SID-Set introduce several core differences and innovations:

Regarding Datasets (SID-Set vs. existing):

• Scale and Diversity: SID-Set is presented as the first dataset of its scale (300K images) specifically designed for social media deepfakes, featuring 100K real, 100K fully synthetic, and 100K tampered images. This addresses the Insufficient Diversity challenge of prior datasets that primarily focused on facial imagery or general scenes with less realistic manipulations.
• Realism: SID-Set utilizes state-of-the-art generative models like FLUX to create synthetic images that are "predominantly indistinguishable from genuine ones through mere visual inspection," overcoming the outdated generative techniques limitation of datasets like GenImage and AIGCD.
• Comprehensiveness: It includes both fully synthetic and tampered images with localized manipulations, crucial for real-world social media content. This contrasts with datasets that emphasize only one type or lack fine-grained annotations.
• Explanation Annotations: SID-Set includes textual descriptions explaining the model's judgment basis for 3,000 images, a novel feature that directly supports the explanation task of SIDA and addresses the limited comprehensiveness of existing datasets which often focus only on binary classification or localization masks.

Regarding Methodology (SIDA vs. existing detection/localization models):

• Multimodal Integration for Multi-task Solution: SIDA leverages the strengths of large multimodal models (VLMs) to perform detection, localization, and explanation simultaneously. Most existing methods are limited to one or two of these tasks (e.g., AntifakePrompt for detection, ForgeryGPT for detection/localization, FakeShield for localization/explanation). SIDA offers a unified framework.

• Specialized VLM Adaptation: SIDA expands the vocabulary of base VLMs (like LISA) with $<DET>$ and $<SEG>$ tokens, specifically designed to extract deepfake detection and segmentation information. It also incorporates a Multihead Attention module to enhance feature interaction between detection and segmentation features, which is crucial for precise localization. This specialized adaptation for deepfake tasks is a key innovation compared to general-purpose VLMs.

• Interpretability: By providing textual explanations of its judgment criteria, SIDA addresses a critical gap in deepfake detection, moving beyond black-box predictions. This enhances trust and utility for users, making the system more transparent.

• Enhanced Localization: The paper claims SIDA achieves superior localization performance, even compared to dedicated IFDL (Image Forgery Detection and Localization) methods and general segmentation VLMs like LISA, by virtue of its specialized design for subtle manipulations.

  In essence, SIDA and SID-Set aim to set a new standard for holistic, robust, and interpretable deepfake analysis in the increasingly complex social media landscape.

4. Methodology

4.1. Principles

The core idea behind SIDA is to leverage the exceptional capabilities of large vision-language models (VLMs) to perform a multi-faceted analysis of social media images for deepfake detection. Traditional deepfake detection often focuses solely on binary classification (real/fake) or localization (identifying tampered regions). SIDA expands on this by integrating detection, localization, and explanation into a unified framework.

The theoretical basis or intuition is that VLMs, having learned strong alignment between visual and textual information, can be adapted to understand subtle visual cues indicative of deepfakes and simultaneously generate natural language explanations for their judgments. By introducing specific tokens ($<DET>$ for detection and $<SEG>$ for segmentation) into the VLM's vocabulary and training it on a diverse, high-quality dataset (SID-Set) with corresponding annotations, the model can learn to:

1. Discern Authenticity: Classify an image as real, fully synthetic, or tampered.

2. Delineate Manipulations: Identify and segment the exact manipulated regions within a tampered image.

3. Provide Interpretability: Generate human-readable textual explanations for why an image is deemed fake or where it's tampered.

   The architecture is inspired by existing VLM designs but is specifically tailored for deepfake analysis, including specialized heads for detection and segmentation, and a mechanism to enhance feature interaction between these tasks.

4.2. Core Methodology In-depth (Layer by Layer)

SIDA's methodology is built upon a VLM base, specifically LISA [29], which already possesses strong multimodal understanding and segmentation capabilities. The authors extend this by introducing specialized tokens and an enhanced feature interaction mechanism for deepfake tasks.

4.2.1. Architecture

The pipeline of SIDA involves processing an image and a textual prompt through a VLM to produce detection results, segmentation masks, and textual explanations.

1. Input Processing: The framework takes an input image, denoted as $x_i$, and a corresponding text prompt, denoted as $x_t$. An example text prompt provided is: "Can you identify if this image is real, fully synthetic, or tampered? Please mask the tampered object/part if it is tampered." These inputs are fed into the VLM.

2. VLM Output and Token Extraction: The VLM processes these inputs. Its primary output is a text description, denoted as $\hat{y}_{\mathrm{des}}$. Crucially, the VLM's vocabulary is expanded with two new tokens: $<DET>$ (for detection) and $<SEG>$ (for segmentation). The hidden layer output of the VLM, denoted as $h_{\mathrm{hid}}$, contains representations for these tokens. This initial processing can be formulated as: $$\hat{y}_{\mathrm{des}} = \mathrm{VLM}(x_i, x_t)$$ Where:

   • $x_i$: The input image.
   • $x_t$: The input text prompt.
   • $\mathrm{VLM}(\cdot)$: The large vision-language model.
   • $\hat{y}_{\mathrm{des}}$: The generated textual description or explanation by the VLM.

3. Deepfake Detection: To perform detection, the representation corresponding to the $<DET>$ token is extracted from the last hidden layer $h_{\mathrm{hid}}$. This extracted representation is denoted as $h_{\mathrm{det}}$. This $h_{\mathrm{det}}$ is then passed through a detection head, $F_{\mathrm{det}}$, which is typically a neural network layer (e.g., a fully connected layer with a softmax activation), to classify the image. The detection head determines if the image is real, fully synthetic, or tampered. The final detection result is denoted by $\hat{\mathrm{D}}$. $$\hat{\mathrm{D}} = F_{\mathrm{det}}(h_{\mathrm{det}})$$ Where:

   • $h_{\mathrm{det}}$: The representation extracted from the $<DET>$ token in the VLM's hidden layer.
   • $F_{\mathrm{det}}(\cdot)$: The detection head, which processes $h_{\mathrm{det}}$.
   • $\hat{\mathrm{D}}$: The predicted detection result (e.g., 'real', 'fully synthetic', 'tampered').

4. Tampered Region Localization (if detected as tampered): If the detection result $\hat{\mathrm{D}}$ indicates that the image has been tampered, SIDA proceeds to predict a segmentation mask for the manipulated regions. The representation corresponding to the $<SEG>$ token, denoted as $h_{\mathrm{seg}}$, is extracted from the hidden layer $h_{\mathrm{hid}}$. To enhance the mask prediction, the representation $h_{\mathrm{det}}$ (from the $<DET>$ token) is transformed using a fully connected layer $F$ to align its dimensions with $h_{\mathrm{seg}}$. This transformed detection feature is then used as a query in a Multihead Attention (MSA) mechanism, with the segmentation feature $h_{\mathrm{seg}}$ serving as both key and value. This step allows the model to capture the relationship between the overall detection decision and the specific regions to be segmented, refining the segmentation features. A residual connection is applied to combine the original and transformed information. $$\begin{array}{rcl} \tilde{h}_{\mathrm{det}} & = & F(h_{\mathrm{det}}) \\ \tilde{h}_{\mathrm{seg}} & = & \mathtt{MSA}(\tilde{h}_{\mathrm{det}}, h_{\mathrm{seg}}) \\ \tilde{h}_{\mathrm{seg}} & = & \tilde{h}_{\mathrm{seg}} + h_{\mathrm{seg}} \end{array}$$ Where:

   • $F(\cdot)$: A fully connected layer that transforms $h_{\mathrm{det}}$.

   • $\tilde{h}_{\mathrm{det}}$: The dimension-aligned representation of the detection feature.

   • $\mathtt{MSA}(\cdot)$: The Multihead Attention mechanism. Here, $\tilde{h}_{\mathrm{det}}$ acts as the query, and $h_{\mathrm{seg}}$ acts as both key and value.

   • $h_{\mathrm{seg}}$: The representation extracted from the $<SEG>$ token.

   • $\tilde{h}_{\mathrm{seg}}$: The enhanced segmentation embedding after attention and residual connection.

     The overall framework is illustrated in the following figure:

     该图像是示意图，展示了SIDA（社交媒体图像检测、定位与解释助手）的工作流程。输入图像和文本后，模型通过最后一层隐藏状态检测图像是否被篡改。若检测结果为篡改，模型会生成篡改区域的掩码，并提供有关篡改位置的详细说明，如“面部的篡改”。该图描述了深度学习模型如何识别和定位图像中的伪造部分。

   Figure 5. The pipeline of SIDA: Given an image x _ { i } and the corresponding text input x _ { t } , the last hidden layer for the ${ \mathrm { < D E T > } }$ token provides the detection result. If the detection result indicates a tampered image, SIDA extracts the ${ < } \mathrm { S E } { \mathrm { G } } { \mathrm { > } }$ token to generate masks for the tampered regions. This figure shows an example where the man's face has been manipulated.

5. Final Mask Generation: Finally, a frozen image encoder, $F_{\mathrm{enc}}$, extracts visual features, $f$, directly from the input image $x_i$. This ensures that the low-level visual details are preserved. The enhanced segmentation embedding $\tilde{h}_{\mathrm{seg}}$ is then combined with these visual features $f$ and fed into a decoder, $F_{\mathrm{dec}}$, to produce the final predicted segmentation mask, $\hat{\mathsf{M}}$. $$\begin{array}{rcl} f & = & F_{\mathrm{enc}}(x_i) \\ \hat{\mathsf{M}} & = & F_{\mathrm{dec}}(\tilde{h}_{\mathrm{seg}}, f) \end{array}$$ Where:

   • $F_{\mathrm{enc}}(\cdot)$: A frozen image encoder (e.g., from a pre-trained vision model).
   • $f$: Visual features extracted from the input image $x_i$.
   • $F_{\mathrm{dec}}(\cdot)$: A decoder network that takes the enhanced segmentation embedding and visual features to produce the mask.
   • $\hat{\mathsf{M}}$: The final predicted segmentation mask for tampered regions.

4.2.2. Training

The training of SIDA is conducted in two main phases: an initial end-to-end training for detection and segmentation, followed by a fine-tuning phase for text generation.

1. Initial End-to-End Training: In this phase, SIDA is trained using a combination of detection loss ($\mathcal{L}_{\mathrm{det}}$) and segmentation mask loss ($\mathcal{L}_{\mathrm{mask}}$).

   • Detection Loss: For the detection task (classifying images as real, synthetic, or tampered), CrossEntropy loss ($\mathcal{L}_{\mathrm{CE}}$) is used. This measures the difference between the model's predicted probabilities for each class ($\hat{\mathrm{D}}$) and the true class label ($\mathbb{D}$).
   • Segmentation Mask Loss: For the segmentation task, a weighted combination of Binary Cross-Entropy (BCE) loss ($\mathcal{L}_{\mathrm{BCE}}$) and DICE loss ($\mathcal{L}_{\mathrm{DICE}}$) is employed.
     • $\mathcal{L}_{\mathrm{BCE}}$ measures pixel-wise classification accuracy (whether each pixel belongs to the tampered region or not), comparing the predicted mask ($\hat{\mathbb{M}}$) with the ground truth mask ($\mathbb{M}$).
     • $\mathcal{L}_{\mathrm{DICE}}$ measures the overlap between the predicted and ground truth masks, which is particularly useful for imbalanced segmentation problems (where tampered regions might be small). The combined loss for this stage is formulated as: $$\begin{array}{rcl} \mathcal{L} & = & \lambda_{\mathrm{det}}\mathcal{L}_{\mathrm{det}} + \lambda_{\mathrm{mask}}\mathcal{L}_{\mathrm{mask}} \\ \mathcal{L}_{\mathrm{det}} & = & \mathcal{L}_{\mathrm{CE}}(\hat{\mathrm{D}}, \mathbb{D}) \\ \mathcal{L}_{\mathrm{mask}} & = & \lambda_{\mathrm{bce}}\mathcal{L}_{\mathrm{BCE}}(\hat{\mathbb{M}}, \mathbb{M}) + \lambda_{\mathrm{dice}}\mathcal{L}_{\mathrm{DICE}}(\hat{\mathbb{M}}, \mathbb{M}) \end{array}$$ Where:
   • $\mathcal{L}$: The total loss for the initial training phase.
   • $\lambda_{\mathrm{det}}$: Weighting factor for the detection loss.
   • $\mathcal{L}_{\mathrm{det}}$: The CrossEntropy loss for detection.
   • $\hat{\mathrm{D}}$: The predicted detection result.
   • $\mathbb{D}$: The ground truth detection label.
   • $\lambda_{\mathrm{mask}}$: Weighting factor for the segmentation mask loss.
   • $\mathcal{L}_{\mathrm{mask}}$: The combined segmentation loss.
   • $\lambda_{\mathrm{bce}}$: Weighting factor for the BCE loss component of segmentation.
   • $\mathcal{L}_{\mathrm{BCE}}(\hat{\mathbb{M}}, \mathbb{M})$: Binary Cross-Entropy loss between predicted mask $\hat{\mathbb{M}}$ and ground truth mask $\mathbb{M}$.
   • $\lambda_{\mathrm{dice}}$: Weighting factor for the DICE loss component of segmentation.
   • $\mathcal{L}_{\mathrm{DICE}}(\hat{\mathbb{M}}, \mathbb{M})$: DICE loss between predicted mask $\hat{\mathbb{M}}$ and ground truth mask $\mathbb{M}$.

2. Fine-tuning for Text Generation: After the initial training, the SIDA model undergoes a second fine-tuning phase. This phase focuses specifically on improving the model's ability to generate detailed textual explanations (textural interpretability). For this, detailed textual descriptions ($y_{\mathrm{des}}$) from 3,000 images in the SID-Set are used as ground truth. The CrossEntropy loss is again used, but this time to optimize the text generation component. The final total loss function, combining all three objectives, is: $$\begin{array}{rcl} \mathcal{L}_{\mathrm{txt}} & = & \mathcal{L}_{\mathrm{CE}}(y_{\mathrm{des}}, y_{\mathrm{des}}) \\ \mathcal{L}_{\mathrm{total}} & = & \lambda_{\mathrm{det}}\mathcal{L}_{\mathrm{det}} + \lambda_{\mathrm{mask}}\mathcal{L}_{\mathrm{mask}} + \lambda_{\mathrm{txt}}\mathcal{L}_{\mathrm{txt}} \end{array}$$ Where:

   • $\mathcal{L}_{\mathrm{txt}}$: The CrossEntropy loss for text generation. The formula $\mathcal{L}_{\mathrm{CE}}(y_{\mathrm{des}}, y_{\mathrm{des}})$ implies that the model's generated text is compared against the ground truth text, where $y_{\mathrm{des}}$ on the right side represents the ground truth.
   • $y_{\mathrm{des}}$: The ground truth textual description.
   • $\mathcal{L}_{\mathrm{total}}$: The overall loss function for the model.
   • $\lambda_{\mathrm{txt}}$: Weighting factor for the text generation loss.

4.2.3. Training Data

The primary dataset used for training SIDA is the SID-Set, which comprises 300K images. To further enhance diversity and robustness, the MagicBrush dataset [79] (after filtering low-quality images) is also incorporated. For the text generation fine-tuning phase, detailed descriptions generated by LLMs (specifically GPT-4o [47]) for 3,000 randomly selected images from SID-Set serve as ground truth.

5. Experimental Setup

5.1. Datasets

The experiments primarily use a newly introduced dataset, SID-Set, and also a benchmark dataset, DMimage, for generalization studies.

5.1.1. SID-Set

SID-Set (Social media Image Detection dataSet) is the core benchmark developed in this paper. It aims to provide a comprehensive resource for deepfake detection, localization, and explanation, tailored for social media.

• Scale: 300K images.

• Composition:

  • Real Images: 100K images sourced from OpenImages V7 [43]. These images offer a wide range of scenarios, reflecting real-world diversity.
  • Synthetic Images: 100K images generated using FLUX [43]. FLUX was chosen after human expert review found it produced highly convincing images, indistinguishable from real ones. These images are based on prompts derived from Flickr30k [51] and COCO [33].
  • Tampered Images: 100K images where specific objects or regions have been replaced or altered. The generation process for these images is detailed below.

• Annotations:

  • Detection labels: real, fully synthetic, or tampered.
  • Masks: For tampered images, segmentation masks delineate the altered regions.
  • Textual descriptions: For 3,000 images, GPT-4o generated textual explanations of the judgment basis.

• Characteristics: High quality, broad diversity (various classes, complex scenes), and elevated realism, with subtle alterations (sometimes just dozens of pixels) and natural-looking local manipulations.

  The tampered image generation process is depicted in the paper's Figure 4 and follows four stages:

1. Stage 1: Object Extraction from Caption: For a given COCO image, GPT-4o [47] is used to extract objects from its caption. If a word matches a COCO class, it's identified; otherwise, nouns are retained. This information is stored in an "ImageCaption-Object" JSON file.

2. Stage 2: Mask Generation: Language-SAM [30] is employed to generate segmentation masks for the identified objects. These masks serve as ground truth for training localization.

3. Stage 3: Tampering Dictionaries: Dictionaries are created for full and partial image tampering. COCO classes are used for object replacement (e.g., replacing "dog" with "cat") and attribute modifications (e.g., adding "happy" or "angry" to a "dog").

4. Stage 4: Image Regeneration: Latent Diffusion [56] is used to modify image captions and regenerate images. For example, changing "cat" to "dog" in a caption leads to an object replacement. This process yielded 80,000 object-tampered images and 20,000 partially tampered images.

   The following figure (Figure 2 from the original paper) shows examples from the SID-Set:

   该图像是一些场景和活动的合成图，展示多种日常生活中的情景，包括街头、运动、聚会等。每一张图像都有独特的主题，体现了丰富的人文和自然元素。

Figure 2. SID-Set examples. The 1st row is the synthetic images, while the 2nd row shows tampered images. (Zoom in to view)

The following figure (Figure 4 from the original paper) shows examples of tampered images and their masks:

该图像是一个示意图，展示了多张图像和对应的掩膜，包含鸟、猫、狗、咖啡桌等不同物体的例子，以及它们的背景区域。每个图像旁边展示了特定物体的掩膜，标识出被篡改或合成的区域。

Figure 4. Examples of tampered images. (Zoom in to view)

5.1.2. MagicBrush Dataset

The MagicBrush dataset [79] is incorporated to further enhance diversity during training. It is used after filtering out low-quality images. MagicBrush is a manually annotated dataset for instruction-guided image editing, which aligns well with SIDA's VLM-based approach and its goal of handling tampered images.

5.1.3. DMimage Dataset

The DMimage dataset [8] is used as an external benchmark to evaluate SIDA's generalization capabilities. This dataset focuses on synthetic images generated by diffusion models. It provides a good test for how well SIDA performs on data not necessarily from its primary training distribution but still relevant to AI-generated content.

5.2. Evaluation Metrics

The paper uses a comprehensive set of metrics to evaluate SIDA's performance across its detection and localization tasks.

5.2.1. Detection Metrics

For the detection task, SIDA classifies images into real, fully synthetic, or tampered. The following metrics are used:

1. Accuracy (Acc):

   • Conceptual Definition: Accuracy measures the proportion of total predictions that were correct. It is a straightforward metric that indicates the overall correctness of the model's classifications.
   • Mathematical Formula: $$\text{Accuracy} = \frac{\text{Number of Correct Predictions}}{\text{Total Number of Predictions}} = \frac{\text{TP} + \text{TN}}{\text{TP} + \text{TN} + \text{FP} + \text{FN}}$$
   • Symbol Explanation:
     • $\text{TP}$: True Positives (correctly predicted positive instances).
     • $\text{TN}$: True Negatives (correctly predicted negative instances).
     • $\text{FP}$: False Positives (incorrectly predicted positive instances, i.e., Type I error).
     • $\text{FN}$: False Negatives (incorrectly predicted negative instances, i.e., Type II error).

2. F1 Score:

   • Conceptual Definition: The F1 Score is the harmonic mean of Precision and Recall. It is particularly useful when dealing with imbalanced datasets or when both false positives and false negatives are costly. It provides a single score that balances the trade-off between precision (how many selected items are relevant) and recall (how many relevant items are selected).
   • Mathematical Formula: $$\text{F1 Score} = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}$$ where: $$\text{Precision} = \frac{\text{TP}}{\text{TP} + \text{FP}}$$ $$\text{Recall} = \frac{\text{TP}}{\text{TP} + \text{FN}}$$
   • Symbol Explanation:
     • $\text{TP}$: True Positives.
     • $\text{FP}$: False Positives.
     • $\text{FN}$: False Negatives.

5.2.2. Localization Metrics

For the localization task, where SIDA predicts segmentation masks for tampered regions, the following metrics are used:

1. Area Under the Curve (AUC):

   • Conceptual Definition: In the context of localization (which can be treated as a binary pixel-wise classification problem for each pixel), AUC typically refers to the Area Under the Receiver Operating Characteristic (ROC) curve. The ROC curve plots the True Positive Rate (TPR) (also known as Recall or Sensitivity) against the False Positive Rate (FPR) (1 - Specificity) at various classification thresholds. AUC measures the overall ability of a binary classifier to discriminate between classes across all possible thresholds. A higher AUC value indicates better performance, where 1.0 is a perfect classifier.
   • Mathematical Formula: $$\text{AUC} = \int_{0}^{1} \text{TPR}(\text{FPR}^{-1}(x)) \, dx$$ where: $$\text{TPR} = \frac{\text{TP}}{\text{TP} + \text{FN}}$$ $$\text{FPR} = \frac{\text{FP}}{\text{FP} + \text{TN}}$$
   • Symbol Explanation:
     • $\text{TPR}$: True Positive Rate.
     • $\text{FPR}$: False Positive Rate.
     • $\text{TP}$: True Positives.
     • $\text{FN}$: False Negatives.
     • $\text{FP}$: False Positives.
     • $\text{TN}$: True Negatives.

2. F1 Score:

   • Conceptual Definition: Similar to its use in detection, the F1 Score for localization evaluates the quality of the predicted segmentation mask by balancing Precision and Recall at the pixel level. It is a robust metric for segmentation, especially for small or irregularly shaped objects.
   • Mathematical Formula: $$\text{F1 Score} = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}$$ The definitions of Precision and Recall are the same as above, but calculated on a pixel-wise basis:
     • $\text{TP}$: Pixels correctly identified as part of the tampered region.
     • $\text{FP}$: Pixels incorrectly identified as part of the tampered region.
     • $\text{FN}$: Pixels from the tampered region that were missed by the prediction.
   • Symbol Explanation:
     • $\text{TP}$: True Positives (pixels).
     • $\text{FP}$: False Positives (pixels).
     • $\text{FN}$: False Negatives (pixels).

3. Intersection over Union (IoU):

   • Conceptual Definition: IoU quantifies the overlap between the predicted segmentation mask and the ground truth mask. It is calculated as the area of their intersection divided by the area of their union. A higher IoU value indicates a better match between the predicted and true masks, signifying more accurate localization.
   • Mathematical Formula: $$\text{IoU} = \frac{\text{Area}(\text{Predicted Mask} \cap \text{Ground Truth Mask})}{\text{Area}(\text{Predicted Mask} \cup \text{Ground Truth Mask})}$$
   • Symbol Explanation:
     • $\text{Area}(\text{Predicted Mask} \cap \text{Ground Truth Mask})$: The number of pixels common to both the predicted and ground truth tampered regions.
     • $\text{Area}(\text{Predicted Mask} \cup \text{Ground Truth Mask})$: The total number of pixels covered by either the predicted or ground truth tampered regions (or both).

5.3. Baselines

5.3.1. Deepfake Detection Models (for SID-Set and DMimage)

SIDA is compared against several state-of-the-art deepfake detection models:

• CnnSpot [17]: A CNN-based method known for its ability to spot CNN-generated images.

• AntifakePrompt [5]: A recent approach that formulates deepfake detection as a visual question answering problem, using prompt-tuned vision-language models.

• FreDect [18]: Leverages frequency analysis for deepfake image recognition.

• Fusing [26]: A method that fuses global and local features for generalized AI-synthesized image detection.

• Gram-Net [37]: Focuses on global texture enhancement for fake face detection.

• UnivFD [46]: Aims to create universal fake image detectors that generalize across various generative models.

• LGrad [59]: Detects GAN-generated images by learning on gradients.

• LNP [2]: A method for detecting generated images using only real images.

  These baselines are representative because they cover a range of approaches, from CNN-based techniques and frequency domain analysis to more recent VLM-based methods and universal detectors, providing a comprehensive comparison for SIDA's detection capabilities. For a fair comparison, these models are evaluated with their original pre-trained weights and also retrained on SID-Set where applicable.

5.3.2. Image Forgery Detection and Localization (IFDL) Models (for SID-Set localization)

For the localization task, SIDA is compared against dedicated IFDL methods and a general-purpose VLM with segmentation capabilities:

• PSCC-Net [36]: Progressive Spatio-Channel Correlation Network for image manipulation detection and localization.

• MVSS-Net [11]: Multi-View Multi-Scale Supervised Networks for image manipulation detection.

• HIFI-Net [21]: Hierarchical Fine-grained Image Forgery Detection and Localization.

• LISA-7B-v1 [29]: A Large Language Model with Reasoning Segmentation capabilities, chosen because SIDA is built upon LISA. This comparison directly evaluates the specialized modifications made in SIDA for deepfake localization.

  These baselines are representative as they include established IFDL methods and a strong VLM foundation model, allowing for an assessment of SIDA's specialized localization performance.

5.4. Implementation Details

• Base Model: LISA is chosen as the base large vision language model due to its strong capability for reasoning-based localization.
• Fine-tuning Strategy: Both LISA-7Bv1 and LISA-13B-v1 (to create SIDA-7B and SIDA-13B) are fine-tuned on the SID-Set using LoRA (Low-Rank Adaptation).
  • LoRA alpha (\alpha): 16
  • Dropout rate: 0.05
• Input Image Size: Input images are resized to $1024 \times 1024$.
• Loss Weights (Eq. 6 for $\mathcal{L}_{\mathrm{total}}$):
  • Detection loss weight (\lambda_{\mathrm{det}}): 1.0
  • Text generation loss weight (\lambda_{\mathrm{txt}}): 1.0
  • Localization loss weight (\lambda_{\mathrm{mask}}): 1.0
• Localization Loss Weights (Eq. 5 for $\mathcal{L}_{\mathrm{mask}}$):
  • BCE loss weight (\lambda_{\mathrm{bce}}): 2.0
  • DICE loss weight (\lambda_{\mathrm{dice}}): 0.5
  • These weights are chosen to maintain a balance between detection and localization.
• Training Stages:
  • During the detection and localization training stage, the image encoder is frozen, and all other modules (VLM, detection head, segmentation head) are trainable.
  • For the text generation stage, only the vision-language models are fine-tuned using the LoRA strategy.
• Optimization:
  • Initial learning rate: $1 \times 10^{-4}$
  • Batch size: 2 per device
  • Gradient accumulation step: 10
• Hardware: Two NVIDIA A100 GPUs (40GB each).
• Training Time:
  • SIDA-7B: 48 hours
  • SIDA-13B: 72 hours

6. Results & Analysis

6.1. Core Results Analysis

6.1.1. Detection Evaluation on SID-Set

The following are the results from Table 2 of the original paper:

Table 2 compares SIDA with other state-of-the-art deepfake detection models on the SID-Set. The numbers in parentheses for baseline methods (e.g., $(↑24.1)$) indicate the performance improvement after retraining on SID-Set, implying the base numbers are from original pre-trained weights.

Key Observations:

• Superior Overall Performance: Both SIDA-7B and SIDA-13B achieve the highest Overall Accuracy and F1 scores (93.5% / 93.5% for SIDA-7B and 93.6% / 93.5% for SIDA-13B). This demonstrates SIDA's effectiveness in classifying real, fully synthetic, and tampered images.
• Strong on Fully Synthetic and Tampered: SIDA models consistently achieve very high Accuracy and F1 scores for Fully synthetic images (around 98%) and Tampered images (around 91-92%). This is particularly important for detecting the types of deepfakes prevalent on social media, where tampered content can be subtle.
• Robust on Real Images: While not the absolute highest in Real Acc/F1 (e.g., Fusing has a slightly higher Real F1 at 92.0, but its overall performance is much lower), SIDA maintains strong performance (around 89-91%) on real images, indicating it does not suffer from high false positive rates for authentic content.
• Baseline Performance Variation:

  • After retraining on SID-Set, many baselines show significant improvements (indicated by ↑). For example, AntifakePrompt shows large gains across the board.

  • LGrad shows the highest Accuracy and F1 score on tampered images after retraining (92.3% and 86.1% respectively, although the table lists $6.8(↑92.3)$ and $12.7(↑86.1)$ which is a bit ambiguous in presentation, likely indicating an initial low value with a massive increase). However, the paper notes that this comes at the expense of lower performance in other metrics, and its high recall stems from misclassifying other types as tampered, leading to a high false positive rate. This highlights SIDA's balanced performance.

  • Some baselines like CnnSpot, FreDect, Fusing, Gram-Net, and LNP struggle considerably with Fully synthetic and especially Tampered images, even after retraining, demonstrating the challenges posed by SID-Set's realistic content.

    The results strongly validate that SIDA achieves superior or comparable performance compared to existing state-of-the-art models for deepfake detection, particularly excelling in the challenging tampered image category, which is critical for social media contexts.

6.1.2. Localization Results on SID-Set

The following are the results from Table 3 of the original paper:

Table 3 presents the forgery localization performance on the SID-Set. The asterisk * indicates that the pre-trained model from the original paper was used due to unavailable training code.

Key Observations:

• SIDA's Superiority: SIDA-7B achieves the best performance across all localization metrics: AUC (87.3), F1 (73.9), and IoU (43.8). This indicates that SIDA is highly effective at precisely identifying and segmenting the manipulated regions within tampered images.
• Comparison with Strong Baselines:

  • PSCC-Net, a dedicated IFDL method, performs quite well with an AUC of 82.1 and F1 of 71.3, but SIDA still surpasses it.

  • LISA-7B-v1, the base VLM for SIDA (fine-tuned on SID-Set), shows decent segmentation capabilities with AUC 78.4 and F1 69.1. However, SIDA significantly improves upon LISA by incorporating specialized deepfake-aware components. This confirms the paper's hypothesis that while LISA has strong general segmentation capabilities, it lacks the specific features needed for subtle deepfake manipulations.

  • MVSS-Net and HIFI-Net show lower performance, possibly indicating their limitations on the complex and diverse SID-Set or due to the use of their original pre-trained models.

    The localization results underscore SIDA's strength in identifying fine-grained manipulations, which is crucial for tackling realistic deepfakes in social media.

6.1.3. Robustness Study

The following are the results from Table 4 of the original paper:

Table 4 evaluates the robustness of SIDA against common image perturbations encountered in social media. SIDA-7B's performance is listed as a reference without degradation.

Key Observations:

• Resilience to Degradation: Despite not being explicitly trained on degraded data, SIDA maintains a high level of performance across various perturbations.

• JPEG Compression: With JPEG quality levels 70 and 80, detection Accuracy and F1 drop by only 4-5 points, and localization AUC, F1, and IoU see similar minor drops. This is a crucial finding as JPEG compression is ubiquitous on social media.

• Resizing: Scaling factors of 0.5 and 0.75 also show minimal impact, with detection and localization metrics remaining very close to the baseline.

• Gaussian Noise: Even with Gaussian noise (variances 5 and 10), performance remains strong, with detection Accuracy and F1 staying above 86% and 89% respectively, and localization AUC remaining above 84%.

  This robustness study highlights SIDA's practical applicability in real-world social media scenarios, where images often undergo various forms of low-level distortion. The model's stable performance under these conditions is a significant advantage.

6.1.4. Test on Other Benchmark (DMimage)

The following are the results from Table 5 of the original paper:

Table 5 evaluates SIDA-7B's generalization capabilities on the DMimage dataset [8], which primarily contains synthetic images generated by diffusion models. For baselines, their original pre-trained weights and hyperparameters were used.

Key Observations:

• Superior Generalization: SIDA-7B achieves the highest Overall Accuracy (91.8%) and F1 score (92.4%) on the DMimage dataset, demonstrating excellent generalization.

• Strong Performance on Fake Images: SIDA-7B has a Fake Accuracy of 90.7% and Fake F1 of 91.0%, significantly outperforming most baselines, particularly those that struggled with synthetic content in Table 2.

• Comparison with AntifakePrompt: AntifakePrompt, another VLM-based method, performs strongly (90.6% Overall Acc), but SIDA-7B still shows a slight edge. This suggests that SIDA's specialized architecture and training on the diverse SID-Set provide a generalization advantage.

• Baseline Struggles: Many CNN-based baselines like CNNSpot, Fusing, and UnivFD show very low Fake Accuracy (e.g., Fusing at 15.5%, CNNSpot at 28.4%), leading to poor Overall Accuracy. This highlights the difficulty these models have in generalizing to diffusion-generated images when not explicitly trained on them or similar data.

  The results confirm that SIDA's design and training enable it to adapt effectively to unseen generative models and datasets, a crucial aspect for practical deepfake detection.

6.2. Ablation Studies / Parameter Analysis

6.2.1. Attention Module

The following are the results from Table 6 of the original paper:

Table 6 presents an ablation study on the attention module within SIDA. The variants are:

• FC: Replaces the Multihead Attention module with fully connected (FC) layers.
• w/o Attention: Removes the attention module entirely.
• SIDA: The full SIDA model.

Key Observations:

• Critical Role of Attention: The full SIDA model (with Multihead Attention) significantly outperforms both ablated variants across all detection and localization metrics.

• Performance Drop without Attention:

  • Removing attention (w/o Attention) leads to a notable drop in Detection ACC (from 93.5% to 90.3%) and F1 (from 93.5% to 89.9%).
  • The drop is even more pronounced for localization: AUC drops from 87.3 to 84.1, F1 from 73.9 to 71.3, and IoU from 43.8 to 38.8.

• FC Not a Sufficient Replacement: Replacing attention with FC layers (FC) shows similar performance degradation, indicating that simple linear transformations cannot capture the complex interactions learned by attention.

  These results strongly underscore the critical role of the Multihead Attention module in SIDA. It is essential for enhancing feature interaction between detection and segmentation features, leading to improved accuracy in both tasks.

6.2.2. Training Weights

The following are the results from Table 7 of the original paper:

Table 7 shows the impact of different weight configurations for the detection and localization losses during SIDA training. The text generation loss weight ($\lambda_{\mathrm{txt}}$) is not varied in this specific ablation, but it's mentioned that $\lambda_{\mathrm{det}}$, $\lambda_{\mathrm{mask}}$, and $\lambda_{\mathrm{txt}}$ are all set to 1.0 in implementation details for the text fine-tuning stage. For the localization loss, $\lambda_{\mathrm{bce}}$ and $\lambda_{\mathrm{dice}}$ are varied.

Key Observations:

• Optimal Configuration: The configuration $\lambda_{\mathrm{det}}=1.0$, $\lambda_{\mathrm{bce}}=2.0$, $\lambda_{\mathrm{dice}}=0.5$ yields slightly better performance with Acc 93.56 and F1 Score 91.01. This is the configuration chosen for the main experiments.

• Impact of Localization Weights: Increasing the weights for BCE ($\lambda_{\mathrm{bce}}$ from 2.0 to 4.0) and DICE ($\lambda_{\mathrm{dice}}$ from 0.5 to 1.0) slightly decreases both Accuracy and F1 Score (from 93.56 to 93.49 and 91.01 to 90.86, respectively). This suggests that over-emphasizing localization during the initial joint training phase can marginally impact overall detection performance.

  This ablation study confirms that careful tuning of loss weights is necessary to achieve the optimal balance between different tasks and maintain model stability and performance. The chosen weights (1.0 for detection, 2.0 for BCE, 0.5 for DICE) represent a well-optimized configuration for the SIDA framework.

6.3. Qualitative Results

The following figure (Figure 6 from the original paper) provides visual results of SIDA on tampered images:

该图像是示意图，展示了 SIDA 在处理篡改图像时的结果。图中展示了正确（a）和失败（b）示例，其中标记了篡改的对象及其具体类型和位置。该框架可以帮助识别和解释图像中的篡改区域。

Figure 6. Visual results of SIDA on tampered images.

Figure 6 showcases SIDA's visual capabilities, illustrating both successful detection and localization cases ((a) correct examples) and challenging failure cases ((b) failure examples). In the correct examples, SIDA accurately highlights the tampered regions with precise masks and provides textual explanations. For instance, it identifies "tampering of the face" or "tampering of the man's head." The failure cases, however, indicate instances where SIDA either misses subtle manipulations or incorrectly identifies regions, highlighting areas for future improvement. These visualizations provide an intuitive understanding of SIDA's practical output and its current limitations.

7. Conclusion & Reflections

7.1. Conclusion Summary

This paper makes significant strides in the field of deepfake detection for social media by introducing SID-Set and SIDA. SID-Set is a comprehensive, large-scale, and highly realistic dataset comprising 300K images (real, fully synthetic, and tampered) with rich annotations including localization masks and textual explanations. It addresses critical gaps in existing datasets related to diversity, realism, and comprehensiveness for social media contexts. SIDA is a novel VLM-based framework that provides a holistic solution for deepfake detection, localization of tampered regions, and explanation of its judgments. By integrating specialized tokens and an attention mechanism within a VLM architecture, SIDA achieves superior performance on the SID-Set and demonstrates strong generalization capabilities on external benchmarks like DMimage, while also showing robustness against common image perturbations. The ability to offer textual explanations significantly enhances the interpretability and utility of deepfake detection technology.

7.2. Limitations & Future Work

The authors acknowledge several limitations and propose future research directions:

1. Dataset Size: While SID-Set is extensive, the authors recognize that the complexity and sheer volume of real social media content demand an even larger dataset. Expanding SID-Set with more images is a crucial objective.
2. Data Domain: The reliance on FLUX for generating fully synthetic images could potentially lead to data skew, even if FLUX produces high-quality images. This skew might impair SIDA's performance on datasets generated by other methods. Future work will explore integrating a wider variety of generative methods to produce a more diverse and higher-quality set of synthetic images, thus mitigating potential biases.
3. Localization Results: Despite SIDA's strong performance, there is still room for improvement in localization. The paper notes that certain tampered regions are not reliably detected, indicating that the model still struggles with extremely subtle or complex manipulations. This area requires further advancements to enhance the precision and recall of segmentation masks.

7.3. Personal Insights & Critique

This paper presents a highly relevant and timely contribution to the deepfake detection landscape. The SID-Set is a significant step forward, directly addressing the critical need for realistic, diverse, and comprehensively annotated data for social media deepfakes. The inclusion of localized tampering and textual explanations moves the field beyond simple binary classification, offering a more practical and trustworthy solution.

Key Strengths:

• Holistic Approach: The integration of detection, localization, and explanation within a single VLM framework is a powerful and elegant solution. This multi-task capability is essential for real-world applications where users need to know not just if an image is fake, but where and why.
• Interpretability: The textual explanations are a standout feature. As deepfake detection models become more complex, their black-box nature can erode user trust. Providing explanations makes the system more transparent and educational, helping users understand the nuances of deepfake identification.
• Robustness and Generalization: The experiments demonstrating SIDA's resilience to common image degradations and its generalization to other diffusion-generated datasets are highly encouraging for practical deployment.

Potential Issues and Areas for Improvement/Further Research:

• Explanation Quality and Bias: While a novel feature, the textual explanations are generated by GPT-4o. This raises questions about the potential for hallucinations or biases in the explanations themselves. Future work could focus on ensuring the factual accuracy and robustness of these AI-generated explanations, perhaps through human-in-the-loop validation or by grounding them more directly in detectable visual artifacts.

• Computational Cost: Large Multimodal Models are computationally expensive to train and deploy. While LoRA helps with fine-tuning efficiency, real-time social media deepfake detection at scale might still pose significant infrastructure challenges. Exploring more efficient VLM architectures or distillation techniques could be valuable.

• Subtle Manipulations and Adversarial Attacks: The paper acknowledges that some tampered regions are not reliably detected. Deepfakes are constantly evolving, with new techniques emerging to create even more subtle and resilient manipulations. Future research needs to continuously adapt the SID-Set and SIDA to counter adversarial attacks designed to fool detectors.

• Ethical Considerations: Releasing a large, high-quality deepfake dataset (like SID-Set) always carries a dual risk: while beneficial for detection research, it could also be misused to train better deepfake generators. The authors should consider clear ethical guidelines for dataset usage.

• Dynamic Social Media Content: Social media content is highly dynamic, with trends and generative models changing rapidly. The SID-Set, though diverse now, will need continuous updates to remain representative of the latest deepfake trends. A framework for continuous dataset generation and model adaptation could be a future direction.

  Overall, SIDA and SID-Set represent a significant advancement, offering a powerful, interpretable, and adaptable solution that pushes the boundaries of deepfake detection for social media. Its insights on multimodal integration and comprehensive analysis could be transferred to other domains requiring fine-grained visual anomaly detection and explanation, such as medical imaging or industrial quality control.